i trained a computer vision model to count chinups roboflow rapid + SAM3 detected the chinup bar - hands - head through natural language prompts then I wrote a python script to count completed reps (head above bar) and annotate the video idea for a gamified fitness app? automated... tracking of your workout, with rep counts per exercise added to your database. this gives progress stats over time, and a leaderboard with your friends these new vision models are neat because they can detect exercise equipment, allowing tracking of dumbbell workouts, bench press, squat, etc.show more

i'll post the code for this demo here in the next couple days: there's 20 computer vision projects there already

today’s experiment is basketball vision:

okay you can stop flexing now 😆 though if i could do 14 chinups i'd probably post about it too!

to train models we must first train ourselves (or some bs like that) lol thank you Max

if i had to do one push up for every 10K tokens consumed i'd be shredded, or maybe dead

he stopped the video at 14 chinups but I heard the man kept on going for another hour

here’s the extended footage

🔥 this is awesome!!! i made a similar computer vision system for overhead presses!! feel free to add it to your fitness app!!

nice !! computer vision is fun :)

this is actually fire for the fitness tech space bro

more exercises and experiments coming soon

LFG

I just finished creating a bouldering app, I’d love to test roboflow rapid + sam3

awesome, landing page looks cool, would be nice to put a short demo video up top are you using mediapipe for body tracking?

thank you!, right now it just very basic and simple, just extracting the frames with ffmpeg and feeding into gpt 5 but ideally i'd love to use media pipe for body tracking for better feedback. looking into using SAM3

awesome yea if you can figure it out with mediapipe it would be much cheaper to run (free)

Great job!! We also have a groundbreaking open-source project based on vision model, if you ever wished your “AI phone” agent actually used the phone for you, check this.

💪

thank you, long live SAM3

the fact it detects the chin-up bar is kind of impressive

i didn’t expect it to work so well for the bar some really great work was done here, congrats to you and the team

you building demos like this is the biggest reward

Idk what’s more impressive, you creating the detection model or actually doing the 14 chin-ups

thank you Ivo, very kind

Cool! Would be interesting to see how it performs after the 13 reps (eg when does it start flagging half reps).

good idea i’ll test that

@roboflow

Cool stuff, if marketed well this could sell pretty good. Personally, I hate taking notes about reps count, automating it is a nice feature.

i’ll give it a try 🫡

🤞🤞

SAM was made exactly for you. Zuckerberg needs to spend a 100 million package on you. 😂Like seriously this is awesome 🔥

my bank account is ready thank you Clive

Can’t even tag Zucky here 😆

Working on the same thing 😁

would be cool to see your version when you have it also using roboflow and sam3, or a different stack?

Started working on sam3, still work in progress.

wait this is so insanely neat and actually looks incredible. super accurate and reliable as well

you could make it detect when you're half-repping and have it sigh in disappointment emotional damage as a service

lmao i love it i'm also imagining the voice of Arnold berating you to do better

if you imagined it is now unfortunately your duty to make it real

just to confirm, you're using the Roboflow API for inference, right?

yes, using the roboflow api in my python script

@bilawalsidhu kinda like your shot counter :-)

@bilawalsidhu i want to try a basketball model next :)

I think schools need this: automatically count students' physical fitness test movements and score them.

Does it track if u dont fully lock out at the bottom?

@defi_isaac right now it still counts "bad" reps like that, but I could add more rules to prevent that (e.g., head needs to lower to a certain distance from bar, arm needs to reach a certain angle)

@defi_isaac In the military you need to start from a full hang position, legs straight down and hand position faces forward on the bar. Since none of these "rules" exist in your version, who exactly is your user demographic? Because it's won't be gym bros and army grunts...

@defi_isaac i just started this today! many improvements can be made to this experiment

Thanks for the reminder 😂

no more twitter until 10 chinups 🚨

DONE

let’s gooo nice work

Cool project. For dangerous lifts like a barbell squat it might also measure the angles between torso and femurs and provide real time coaching tips.

that’s a good idea could add mediapipe body tracking to this workflow, which can be used to measure joint angles

Does the AI model know you are standing on a chair?

your honor; i swear these were legitimate reps 🙇‍♂️

Yeah! Nice one mate! The home gym just got dynamic!

thank you captain

Bro is creating the ultimate working out app and does not even know it.

I had the same idea in mind The difference was I wanted to create (train) the model from scratch. I think that will be a little hard.

you’ll probably learn a huge amount in that process, go for it

great dude

What a nice project you did my friend, congratulations!!!

The biggest barrier in my opinion to this kind of application is the fact that you need to set up your phone to record video the whole time you are working out. That is a hassle that I personally would not do.

yea that’s a fair point for me it’s ok for a home workout where i’m staying near the same place. just leave the phone running in one spot but it’s too much effort if i’m going to be moving around a gym

Yes. It's awkward to set up a camera in public.

this was me after i stopped the video

Already exists kinda but def can be improved upon

dayum. cool tech, cooler dude

Your doing great, and starring in your own IRL demos

much appreciated Greg, more IRL demos coming soon

I have never used Roboflow Rapid, but does it mean you use an API from Roboflow for the SAM 3 model?

yes, but the api does more than that i believe

What about the cost? Since running SAM 2 on a long video requires a big GPU machine.